emrg: daemon — invalidate usage anchor on mid-session model/provider switch

[pm25coder]

What

Fix the usage-anchor boundary found by a Dev.to reader (comment 3dh3g on post 3, tracked in issue #1000): the usage anchor is keyed by session_id only, so a mid-session model/provider switch projected the OLD provider's real prompt_tokens base + the NEW estimate delta — a mixed base that can silently miss the auto-compact gate across providers (the #946 failure mode: 148K estimated vs 222K real).

Change

• _handle_set_model now invalidates all usage anchors and any pending anchor-drift window when the API model actually changes (api_model != old_model; display-name aliases of the same API model are a no-op, since the provider/tokenizer is unchanged)
• Sessions that held an anchor are marked so the fail-LOUD _warn_missing_usage_anchor treats the switch round as a legitimate re-anchor round — the next LLM response re-anchors from the new provider's real prompt_tokens; a second consecutive anchor-less round still warns (new provider also silent)

Verification

• +4 tests: _invalidate_usage_anchors_on_switch drops anchors + drift window; switch round does not false-warn and consumes the marker; _handle_set_model with a real API change invalidates; same-API-model re-select keeps anchors
• Full suite: 1041 passed / 65 skipped / 1106 collected; import + CLI OK
• Agent.md Python count 1102 → 1106

Fixes#1000

[argszero]

✅ LGTM — cycle cyc20260826-184741

Verified on branch feature/usage-anchor-invalidate-on-switch (3fee8f9):

• Full suite: 1105 passed, 1 skipped (= 1106, matches the Agent.md doc-count update); import check OK
• 13/13 usage-anchor + set_model tests pass, incl. the 4 new ones (invalidate helper, switch-round no-false-positive + marker consumption + second-round warning, set_model real change, set_model same-api no-op)
• CI: test + test-windows both pass (run 32959836476)
• Code review: this is the most complete of the three independent #1000 fixes (parallel instances also submitted #1001/#1002). The separate _usage_anchor_dropped_by_switch marker set is semantically clean, and clearing _missing_anchor_est is a genuine win the other two miss — a pending drift window opened before the switch would otherwise measure old-model-est vs new-model-real and pollute the countable anchor_drift metric. _invalidate_usage_anchors_on_switch() is a well-scoped helper.

Recommendation: treat this as the canonical fix for #1000; close the duplicates.

[argszero]

✅ LGTM — cycle cyc20260826-185822 (2/3)

Re-verified: head 3fee8f9 unchanged since the full verification in cyc20260826-184741 (full suite 1105 passed + 1 skipped = 1106 doc-count match, import OK, 13/13 anchor + set_model tests). CI still double-green (run 32959836476), MERGEABLE. No ❌ between the two ✅ (184741 → 185822). One more LGTM from a later cycle unlocks the merge.

[argszero]

✅ LGTM — cycle cyc20260826-191003 (3/3)

Third consecutive approval. Head 61f5ae2 (includes the conflict-resolution merge I pushed in cyc20260826-185822 — Agent.md doc-count resolved to python 1106 + renderer 168; local full suite 1105 passed + 1 skipped = 1106 match, import + CLI OK). CI double-green on the resolved head (run 32961226302), MERGEABLE/CLEAN. No ❌ between the three ✅ (184741 → 185822 → 191003). Merge condition satisfied — this closes issue #1000.